Evaluating Impacts of Climate Change on Traditional Mexican Maize

Presented to:
Dr. Russell Weaver (Chair)
Dr. Jennifer Jensen
Dr. Thomas Ballinger

M.S. Thesis Proposal
Department of Geography
Texas State University, San Marcos, TX
github: www.github.com/stevenconnorg/maices-enm
contact: scg67@txstate.edu

1. Model current relationships between maize landraces and their niches
   • Project maize niches into geographic space
   • Forecast these projections with future climate models
2. Identify patterns
   • Evaluate spatio-temporal relationships between indigeneity and maize diversity (Variable importance)
   • Quantify maize landrace diversity and it's projected change (Shannon Diversity Index)
   • Identify similarities among species occurrence patterns (Jaccard similarity)
   • Geovisualization
3. Provide the groundwork for future work

1. How do social and environmental factors contribute to maize landraces' current distribution?
2. What are the characteristics of each maize landraces ''bioclimate envelope''?
3. Can the contributions of indigeneity and society to maize landraces distribution and diversity be discerned and modelled at the national level?
4. What is the nature of the relationship between maize landrace diversity and indigeneity in Mexico?
5. How is climate change projected to alter current maize landrace distributions?
6. Which maize landraces are suseptible of extinction/expansion?

• Uses species presence and/or absence observation data and ecological data at those observation sites to estimate complex relationships of species niche (Elith et al. 2011).
• Project hyperdimensional ecological space onto 2- or 3D geographic space
• Numerous uses including:
  1. Identifying potential spread of vector borne diseases
  2. Identifying range of invasive species
  3. Threatened species management
  4. Crop modelling
• Numerous ecological niche modeling algorithms have been developed to project ecological niche spaces into geographic spaces (e.g.: Bioclimatic Envelopes, Genetic Algorithm Rule-set Production, Maximum Entropy, Artificial Neural Networks, Support Vector Machines, Boosted Regression Trees)
• Ensemble Ecological Niche Modelling weights individual models based on performance; increases performance (Filho et. al, 2010, Araujo & New 2007)
• WA and Mean consensus methods provided significantly more robust predictions than single-models and the other consensus methods (Marimon et al., 2009)

1. Georeferenced maize observations
2. Ethno-linguistic Data
3. Gridded Climatologies
   • derived “Bioclimatic Variables”
4. Topography/Land Cover

• WorldClim

  • 30 arc-second resolution
  • ANUSPLIN up to 60,000 weather stations (with >10 years data)
  • largest available archive of downscaled climate data for ecology (Guisan et al., 2017)
  • >2,700 citations for ecological studies including ENMs and SDMs (Booth et al., 2014)

• Current Data from WorldClim 2.0 (Fick & Hijmans, 2017)

  • 1970-2000
  • Maximum and Minimum Monthly Temperatures (°C * 10)
  • Monthly Precipitation (mm)
  • also includes solar radiation, windspeed, water vapor pressure

• Future Data from Worldclim 1.4 (Hijmans et al., 2005)

  • 2041-2060, 2061-2080
  • dynamically-downscaled from 19 GCM used in CMIP5 (Coupled Model Intercomparison Project 5; Meehl et al., 2009) and the Fifth Assessment Report (IPCC, 2013 )
  • RCP 8.5

• Ensemble future climate models to reduce uncertainty across:
  • Relative Concentration Pathways
  • GCM
• Monthly future climatologies were averaged across five GCMs:
  • CCSM4 (Community Climate System Model, UCAR)
  • MIROC5 (Model for Interdisciplinary Research on Climate)
  • MPI-ESM-LR (Max-Plank Institute)
  • HADGEM2-ES (Met Office Hadley)
  • GFDL-CM3 (Geophysical Fluid Dynamics Laboratory )
• Any of these five model performs better singularly than ensemble of all other models in predicting previous climatologies (Conde et al. 2011)
• 5th National Communication of Mexico for the United Nations Framework Convention on Climate Change (2012)
• Here, using updated analogs of GCMs used in (Conde et al. 2011)

• Derived from first ecological niche modelling alogorithm 'BIOCLIM' (Nix, 1986)
• More ecologically importance predictor variables
• Can reduce multicollineary, VIF; expands options
• Calculated from monthly Tmax, Tmin, Tmean, and Prec
• 19 Bioclimatic proxy variables from monthly means using 'dismo' (Hijmans et al., 2017)
• Potentially 16 'complementary' variables from 'envirem' package in R (Title & Bemmels, 2017)
  • Must use (WorldClim 2.0) current terrestial solar radiation for both current and future climatologies
  • Must average Tmin and Tmax to get Tmean (.99 R²⁾ (Title & Bemmels, 2017)

FAO Harmonized Soil Database 1.2 (Fischer et al., 2008)

• Soil Quality Data (Ordered Factor 1-4; 'No or slight limitations' to 'Very severe limitations')
  1. Nutrient availability
  2. Nutrient retention capacity
  3. Rooting conditions
  4. Oxygen availability to roots
  5. Excess salts
  6. Toxicity
  7. Workability (constraining field management)
• Land cover (Continuous)
  1. rain-fed cultivated land
  2. irrigated cultivated land, according to GMIA 4.0
  3. total cultivated land
  4. forest land, calibrated to FRA2000 land statistics
  5. grass/scrub/woodland
  6. built-up land (residential and infrastructure)
  7. barren/very sparsely vegetated land
  8. Mapped Water Bodies

1990 1^st - 4^th major indigenous languages at municipio level

• Indigenous Population Percentage
• Land Cover Classification
• Soil Quality 1-7
• 19 'dismo' bioclimatic variables
• 16 supplemental 'envirem' bioclimatic variables

Method

• 'usdm' R package vif function (Naimi et al., 2014)
• preferred over correlation thresholds due to 'hidden' correlation structures (Guisan et al., 2017)
• correlation structures may change over time
• select variables of ecological importance
• 'vifstep' with VIF threshold of 10 to get remaining variables (Guisan et al., 2017)

               Variables          VIF
1        BIO1AnnMeanTemp 1.261212e+05
2   BIO2MeanDiurnalRange 3.433839e+02
3           BIO3Isotherm 1.008546e+02
4              BIO4TSeas 6.573645e+03
5      BIO5TWarmestMonth          Inf
6   BIO6MinTColdestMonth          Inf
7          BIO7TAnnRange          Inf
8      BIO8MeanTWettestQ 5.934433e+01
9       BIO9MeanTDriestQ 7.722524e+01
10    BIO10MeanTWarmestQ 2.508201e+03
11    BIO11MeanTColdestQ 1.265249e+04
12          BIO12AnnPrec 4.731866e+02
13 BIO13PrecWettestMonth 9.654969e+01
14  BIO14PrecDriestMonth 7.876804e+01
15     BIO15PrecSeas.COV 1.770142e+01
16     BIO16PrecWettestQ 1.680402e+02
17      BIO17PrecDriestQ 1.399311e+02
18     BIO18PrecWarmestQ 8.106017e+00
19     BIO19PrecColdestQ 2.657508e+01
20             EVMannPET 1.820028e+02
21     EVMthornthwaiteAI 1.348320e+08
22         EVMclimaticMI 1.348336e+08
23     EVMcontinentality 2.263212e+03
24          EVMembergerQ 7.249427e+01
25    EVMgrowingDegDays0 1.255414e+05
26    EVMgrowingDegDays5 1.173782e+02
27     EVMmaxTempColdest 5.476294e+02
28     EVMminTempWarmest 3.243490e+02
29    EVMmonthCountByT10 4.614369e+00
30        EVMPETColdestQ 2.230254e+02
31         EVMPETDriestQ 6.754274e+01
32            EVMPETseas 4.166684e+02
33        EVMPETWarmestQ 1.473191e+02
34        EVMPETWettestQ 5.270414e+01
35    EVMthermicityIndex 6.416108e+01
36             TotalCult 5.655149e+03
37               IrrCult 1.057087e+03
38          Rain.fedCult 4.608966e+03
39              Forested 4.483750e+01
40        Grass.Woodland 4.050543e+01
41                Barren 1.699018e+01
42                 Urban 1.927545e+00
43                 Water 2.626440e+00
44             roughness 1.458907e+01
45                aspect 1.029889e+00
46                 slope 1.279400e+01

presvifstep@results

               Variables      VIF
1      BIO8MeanTWettestQ 3.169749
2  BIO13PrecWettestMonth 7.203552
3      BIO15PrecSeas.COV 5.426713
4      BIO18PrecWarmestQ 5.112587
5      BIO19PrecColdestQ 6.479971
6      EVMminTempWarmest 3.258270
7     EVMmonthCountByT10 2.206976
8         EVMPETColdestQ 5.539127
9          EVMPETDriestQ 2.507005
10        EVMPETWarmestQ 4.670706
11               IrrCult 1.169090
12          Rain.fedCult 1.599255
13        Grass.Woodland 2.314950
14                Barren 1.730909
15                 Urban 1.075483
16                 Water 1.107406
17                aspect 1.023113
18                 slope 1.446115

presvifcor@results

              Variables       VIF
1  BIO2MeanDiurnalRange 18.522678
2          BIO3Isotherm 23.926854
3     BIO8MeanTWettestQ  3.749390
4      BIO9MeanTDriestQ 23.234599
5     BIO15PrecSeas.COV  6.459824
6     BIO18PrecWarmestQ  3.463208
7     BIO19PrecColdestQ  7.060739
8             EVMannPET 55.203847
9     EVMminTempWarmest 28.195976
10   EVMmonthCountByT10  2.540841
11        EVMPETDriestQ 15.802205
12       EVMPETWarmestQ 45.516679
13   EVMthermicityIndex 13.293752
14              IrrCult  4.539498
15         Rain.fedCult 14.357032
16             Forested 45.296046
17       Grass.Woodland 40.844532
18               Barren 15.374231
19                Urban  1.777139
20                Water  2.225376
21               aspect  1.027214
22                slope  1.546577

• Ecological Niche modelling with biomod2 with up to
  • 10 modelling algorithms
    • 'GLM','GBM','GAM','CTA','ANN','SRE','FDA','MARS','RF','MAXENT.Phillips', 'MAXENT.Tsuruoka'
  • 10 evaluation metrics
    • 'KAPPA', 'TSS', 'ROC', 'FAR', 'SR', 'ACCURACY', 'BIAS', 'POD', 'CSI' and 'ETS'
  • n modelling repetitions
  • n PA pseudo-absence repetitions
• Data format tuning with 'BIOMOD_Tuning' function
• Ensemble landrace models by weighted-average by model
• Average weighted-averages across models
• Parallel process on TxState LEAP HPC Cluster with OpenMpi/Rmpi and SLURM
  • 64 nodes with 28 CPUs, up to 128 Gb RAM

Barbet-Massin et al. 2012

1. ~ 10,000 of pseudo-absences when using regression techniques
2. ~ 10 runs with ~100 pseudo-absences with multiple adaptive regression splines and discriminant analyses;
3. Equal number of pseudo-absences/presences with ~10 runs for classification techniques (e.g.: BRT, CART and RF).
4. Random selection of pseudo-absences when using regression techniques
5. Environmentally stratified pseudo-absences when using classification and machine-learning techniques

• Binary distribution maps using probability thereshold
• Alpha-Diversity and Range Change/Extinction Rate
• Diversity Indices (and change)
• Evaluation across model type, maize landrace
• Cluster Analysis of Variable Importance
• Evaluation Importance of Indigeneity
• Dendrogram of distribution similarities ('fuzzySim' in R)
• Geovisulation of Impacted Indigenous Communities

• Averaging good-performing updated GCMs for forcasting
• Ensemble modelling
• Expanded datasets
• Careful pseudo-absence selection
• WorldClim 2.0 > 1.4
• CMIP5 Data